v0.2.2
MinishLab/model2vecv0.2.2Oct 3, 2024by Pringled
AI Summary
Bug fixes targeting tokenizer errors, attention masks, and device selection mechanisms.
Key Highlights
- Added explicit errors for BPE and unigram tokenizers
- Fixed attention mask crash during distillation
- Allowed PCA dims to equal model dims
- Added device selection mechanism
New Features
- Device selection mechanism
Full Release Notes
## What's Changed * docs: Update results by @Pringled in https://github.com/MinishLab/model2vec/pull/51 * fix: convert bfloat16 to float because of numpy incompatibility by @stephantul in https://github.com/MinishLab/model2vec/pull/53 * fix: Add explicit errors for BPE and unigram, return tokenizer without cha… by @stephantul in https://github.com/MinishLab/model2vec/pull/54 * Add from model by @stephantul in https://github.com/MinishLab/model2vec/pull/57 * fix: Attention mask being None crashes distillation by @stephantul in https://github.com/MinishLab/model2vec/pull/58 * fix: allow PCA dims == dims by @stephantul in https://github.com/MinishLab/model2vec/pull/59 * enh: Add device selection mechanism by @stephantul in https://github.com/MinishLab/model2vec/pull/60 * Fix in README by @stephantul in https://github.com/MinishLab/model2vec/pull/61 * Bumped version by @Pringled in https://github.com/MinishLab/model2vec/pull/62 **Full Changelog**: https://github.com/MinishLab/model2vec/compare/v0.2.1...v0.2.2